Semantic search using a single-source semantic model

Abstract

Techniques for providing semantic search of a data store are disclosed. A similarity metric of a document comprising the data store to a concept represented in a semantic model derived at least in part from a reference source that includes content not included in the data store is determined. A relevance metric of a search query to the concept is computed. The similarity metric and the relevance metric are used to determine, at least in part, a ranking of the document with respect to the search query.

Description

BACKGROUND OF THE INVENTION

In the field of information retrieval, semantic search techniques have been used to build a semantic model from a set of documents (webpages, emails, or documents on a file system, for example), and given a search query, find the set of documents that best relate to that query. The conventional method has been to build an inverted index of all words in a document across all documents, and then using various relevancy metrics, compare the words of the search query (assumed to be another kind of document) against the index, and finding a ranked set of files that are ‘closest’ to the query. In practice, this serves to simulate semantic search because words that represent a semantic concept tend to cluster together in co-occurrences.

Early methods involved techniques such as Latent Semantic Analysis to calculate the singular value decomposition (SVD) of a matrix derived from the inverted document-word index. For reasonable accuracy, one must specify the k number of dimensions in the Eigen decomposition ahead of time, which can dramatically affect overall search results. More recent approaches, based on principled probabilistic models that bypass the resource intensive SVD computation, including probabilistic latent semantic analysis (PLSA) and ranking support vector machines (SVMs), accomplish the same task, but also require that the number of concepts be known at training time.

While this works quite well in limited domains (such as spam-mail filtering), it proves infeasible for full-fledged desktop search. This is typically due to: a) the user has few or no files to create an index, leading to sparsity of data, and therefore sub-optimal searches; b) the user has hundreds of gigabytes of data, leading to huge indexes and inexcusable computation times when building the model; or c) even with a reasonable index size, it can be difficult finding the optimal k parameter for each individual dataset.

SUMMARY

In accordance with some implementations, a method is provided for semantic searching using a single-source semantic model. The method may be performed at an electronic device with one or more processors and memory. The method includes determining a similarity metric of a document comprising a data store to a concept represented in a semantic model. The semantic model is derived at least in part from a reference source that includes content not included in the data store. The method further includes computing a relevance metric of a search query to the concept. For example, the semantic model is used to compute a degree to which the query terms are relevant to one or more concepts comprising the semantic model. The method further includes using the similarity metric and the relevance metric to determine, at least in part, a ranking of the document with respect to the search query.

In some implementations, the method includes using the reference source to build the semantic model.

In some implementations, the reference source comprises a body of content containing items representing a wide range of concepts.

In some implementations, determining the similarity metric includes embedding the document in the semantic model. In some implementations, computing the relevance metric includes embedding the search query in the semantic model.

In accordance with some implementations, a system including a processor is provided. The processor is configured to determine a similarity metric of a document comprising a data store to a concept represented in a semantic model. The semantic model is derived at least in part from a reference source that includes content not included in the data store. The processor is further configured to compute a relevance metric of a search query to the concept. For example, the semantic model is used to compute a degree to which the query terms are relevant to one or more concepts comprising the semantic model. The processor is further configured to use the similarity metric and the relevance metric to determine, at least in part, a ranking of the document with respect to the search query.

In accordance with some implementations, a computer program product is provided, the computer program product being embodied in a tangible, non-transitory computer readable storage medium. The computer program product includes instructions for determining a similarity metric of a document comprising a data store to a concept represented in a semantic model. The semantic model is derived at least in part from a reference source that includes content not included in the data store. The computer program product further includes instructions for computing a relevance metric of a search query to the concept. For example, the semantic model is used to compute a degree to which the query terms are relevant to one or more concepts comprising the semantic model. The computer program product further includes instructions for using the similarity metric and the relevance metric to determine, at least in part, a ranking of the document with respect to the search query.

BRIEF DESCRIPTION OF THE DRAWINGS

Various embodiments of the invention are disclosed in the following detailed description and the accompanying drawings.

FIG. 1 is a flow chart illustrating an embodiment of a process to search a data store.

FIG. 2 is a block diagram illustrating an embodiment of a system to perform semantic search.

FIG. 3 is a flow diagram illustrating an embodiment of a process to build a semantic model.

FIG. 4 is a block diagram illustrating an embodiment of a semantic model.

DETAILED DESCRIPTION

The invention can be implemented in numerous ways, including as a process; an apparatus; a system; a composition of matter; a computer program product embodied on a computer readable storage medium; and/or a processor, such as a processor configured to execute instructions stored on and/or provided by a memory coupled to the processor. In this specification, these implementations, or any other form that the invention may take, may be referred to as techniques. In general, the order of the steps of disclosed processes may be altered within the scope of the invention. Unless stated otherwise, a component such as a processor or a memory described as being configured to perform a task may be implemented as a general component that is temporarily configured to perform the task at a given time or a specific component that is manufactured to perform the task. As used herein, the term ‘processor’ refers to one or more devices, circuits, and/or processing cores configured to process data, such as computer program instructions.

A detailed description of one or more embodiments of the invention is provided below along with accompanying figures that illustrate the principles of the invention. The invention is described in connection with such embodiments, but the invention is not limited to any embodiment. The scope of the invention is limited only by the claims and the invention encompasses numerous alternatives, modifications and equivalents. Numerous specific details are set forth in the following description in order to provide a thorough understanding of the invention. These details are provided for the purpose of example and the invention may be practiced according to the claims without some or all of these specific details. For the purpose of clarity, technical material that is known in the technical fields related to the invention has not been described in detail so that the invention is not unnecessarily obscured.

Providing semantic search using a single-source semantic model across data stores is disclosed. In various embodiments, a reference source of data is used to build a semantic space. Example reference sources include, without limitation, Wikipedia or other sources of rich, modern concepts that are sorted by concept type (Articles) and even sub-concepts of concepts (Headings within Articles). Wikipedia also has the advantage of being available in a number of languages other than English, allowing us to provide multilingual semantic searches. With a single model, the similarities of a user's documents are computed against the semantic model and those values are stored, to be accessed at search time. When the user enters a search query, and after some appropriate pre-processing (such as query expansion, or augmenting the query with related terms), the query is compared against the model to find the c ranked set of concepts relevant to the query, and from those, the user documents that are closely related to the query (i.e., to the same concepts which have been determined to be relevant to the query) are determined.

FIG. 1 is a flow chart illustrating an embodiment of a process to search a data store. For each of one or more content containing items, such as documents or other files, comprising a data store, a similarity metric relative to a semantic model built at least in part using a reference source external to the data store is determined (102). For example, the similarity of each of a plurality of documents or other content containing items to documents, concepts, etc. comprising a semantic model built using a broad-based reference source, such as Wikipedia or another world knowledge store, may be determined. The determination is made in various embodiments prior to a time when a search query is received and processed. Similarity values are computed in some embodiments by embedding documents comprising the data store into the semantic model, e.g., as described more fully below, and the similarity values are stored for use at search query processing time. A search query is received, and a relevance metric of the search query to documents, concepts, etc. comprising the semantic model is computed (104). The similarity metric values associated with the respective content containing items in the data store and the relevance metric computed for the search query are used to identify and rank content containing items within the data store (106). In various embodiments, this approach enables content containing items within the data store that are the most similar to those concepts, documents, etc. within the semantic model that have been determined to be most relevant to the search query that is being processed to be identified and ranked based at least in part on their respective degree of similarity to those concepts, etc.

FIG. 2 is a block diagram illustrating an embodiment of a system to perform semantic search. In the example shown, a semantic model generator system 202, such as a computer, is connected via the Internet 204 to one or more reference source(s) 206, such as the body of articles comprising the Wikipedia in a given language, and/or one or more other reference sources. The semantic model generator system 202 ingests and processes documents comprising the reference source to build a semantic model, for example using latent semantic analysis and/or associated techniques, based on the reference source. The semantic model is store in a semantic model database 208. In the example shown, a user device 210, such as a desktop or portable computer, a mobile smartphone, tablet computer, or other mobile device, is connected to the Internet 204 via a wired and/or wireless connection 212. In various embodiments, a data store comprising and/or otherwise associated with user device 210 is processed. Examples include, without limitation, documents and/or other content items stored on the user device, such as in a file system; application data objects, such as emails, calendar items, contacts, etc.; and user device vendor and/or software provider provided content, such as help articles. In some embodiments, the semantic model generator 202 embeds content containing items from the user device into the semantic model built using the reference source(s) 206 and computes the similarity of each content item to concepts represented in/by the semantic model. In some embodiments, the resulting semantic model (i.e., now incorporating the user device data store content items) is stored on the user device 210 and used at search query processing time to process search queries. In some embodiments, the model is stored in the semantic model database 208 and search queries are processed by accessing the model at runtime via the Internet 204. In various embodiments, the user device 210 is configured to use the semantic model to process search queries, for example by using the semantic model to compute a degree to which the query terms are relevant to one or more concepts comprising the semantic model and in turn using the semantic model to identify and rank content items in the user device's data store that are most similar to the concepts to which the query has been determined to be relevant. Examples of search queries include without limitation desktop (or other device content) search queries, help topic search queries, etc.

FIG. 3 is a flow diagram illustrating an embodiment of a process to build a semantic model. In the example shown, a reference source that encompasses a wide range of human knowledge, such as Wikipedia, or another online reference source, is used to build a semantic model (302). For example, Wikipedia articles and/or portions thereof in some embodiments are processed to populate columns in a term frequency matrix such as those commonly used to perform latent semantic analysis. Content containing items, such as documents, comprising a user device or other data store are embedded into the semantic model (304). For example, documents stored on the user device may be processed similarly to articles comprising the reference source, each being added as a column to the matrix described above. The SVD and/or other transforms may be performed (306) to reduce the dimensionality (i.e., size and complexity) of the matrix while preserving the concept similarity relationships among documents represented in the model. In some embodiments, the value of optimal k will depend on the analysis of the reference data which is used to train the model. One can relate the value of k to the detail in which the user expects the results. A lower k value will provide very product concepts that may overlap, while a higher k may provide too much detail that may compromise the larger context to be determined.

FIG. 4 is a block diagram illustrating an embodiment of a semantic model. In the example shown, the semantic model 400 includes a first portion 402 comprising columns each representing an article or other content containing unit included in a reference source, such as Wikipedia, and a second portion 404 representing documents or other content containing items from a data store, such as file system objects on a user device. Individual portions (e.g., documents) are represented as columns, such as columns 406 and 408 in portion 402 and column 410 in portion 404. (The width of the columns 406, 408, 410 relative to the portions 402, 404 as shown in FIG. 4 is exaggerated for clarity.) While the portions 402 and 404 are shown as being separate contiguous portions of model 400, in some embodiments, documents comprising the data store may be interspersed among articles from the reference source, for example as a result of mathematical transformations performed to facilitate runtime performance and/or as a result of semantic relationships between documents from the data store and articles (or other content units) from the reference source.

In some embodiments, the multi-lingual nature of Internet-based references, such as Wikipedia, allows locale-dependent semantic searches to be enabled. For example, if the local data store is (primarily) in French, one could find the semantic relevance to the models stored using French data stores. In some embodiments, one could use a language identification system to determine the locale for the query or let the user specify the locale as a part of the input string. The locale determined is then used to reference the appropriate semantic model.

The approach disclosed herein enables latent semantic analysis techniques to be used to relate documents comprising a data store—in which only a limited number of concepts, the number and names of which may not be known in advance or easily determined, may be represented explicitly—to concepts represented in the reference source but not explicitly represented in the data store. For example, using the approach described herein user device documents relating to the subject of soccer (for example, 410 in FIG. 4) may be found to be responsive to a search query using terms such as “sports” or even “tennis” based on the previously determined similarity of those documents to “sports” and “tennis” related articles in the reference source (for example 406 and/or 408 of FIG. 4), even if the relationship between the soccer-related documents and the concepts of “sports” generally or other sports such as “tennis” may not have been determined and/or as readily apparent if latent semantic processing techniques had been applied to documents comprising the data store in isolation. In another example, the query “French war movies” may reference war movies included in the local data store, and may include concepts such as “war movies about France” or “French movies about wars” or “war movies in the local data store.”

In various embodiments, techniques disclosed herein are used to provide semantic search capability with respect to a defined body of content, using a semantic model based on content not limited to that defined body of content. Examples include, without limitation, searching a local file system or portion thereof (e.g., directory, subdirectory, folder, etc.); searching for emails (Mail); searching iChat discussions by topic; and searching a user's songs by lyric meaning or concepts expressed in the lyrics (for example, using as a reference model the entire iTunes or another music catalog, but the search would be for songs on a user's device, including potentially songs that might not be known to iTunes).

In various embodiments, the same reference source and resulting base semantic model (prior to embedding user or other content) is used to provide semantic search functionality with respect to the respective data stores of a plurality of distinct users and/or for a single user across multiple data stores, for example across multiple user devices each have a subset of the user's documents or other data containing objects stored thereon. Using the same core semantic model across users and/or devices enables a more uniform semantic search experience to be provided across users and/or devices, since regardless of the quantity and/or nature of the content on a particular user device (or other data store), the content items comprising the data store to be searched at search time are related as disclosed herein to concepts represented in a common semantic model based on the reference source.

In some embodiments, the core semantic model may be updated from time to time, for example to incorporate additions or other change to the reference source, such as articles added to the reference source and describing subject matter not included previously in the reference source, such as a recent major world event or an emerging world leader or other public figure. In some embodiments, the updated core semantic model is used to update (for example, regenerate) the user and/or device specific models generated based thereon. For example, user content, which also evolves over time, may be embedded in the updated core semantic model based on the reference source, and new similarity metrics computed for the user documents or other content containing items.

Although the foregoing embodiments have been described in some detail for purposes of clarity of understanding, the invention is not limited to the details provided. There are many alternative ways of implementing the invention. The disclosed embodiments are illustrative and not restrictive.

Claims

1. A method, comprising:

providing a data store including documents;

providing a semantic model including a plurality of concepts, wherein the semantic model is derived at least in part from a reference source that includes content not included in the data store;

determining at least one similarity metric for each document of the plurality of documents, wherein each respective similarity metric represents a similarity between a respective document of the plurality of documents and a respective concept of the plurality of concepts in the semantic model;

receiving a search query;

computing at least one relevance metric of the search query, wherein each relevance metric represents a relevance of the search query to a respective concept of the plurality of concepts represented in the semantic model; and

determining a ranking of at least a subset of the plurality of documents with respect to the search query using at least the at least one similarity metric and the at least one relevance metric.

2. The method of claim 1, further comprising using the reference source to build the semantic model.

3. The method of claim 1, further comprising using the semantic model to provide semantic search functionality with respect to a plurality of data stores.

4. The method of claim 1, wherein the reference source comprises a body of content containing items representing a wide range of concepts.

5. The method of claim 1, wherein the reference source comprises an online source of articles on a wide range of subjects.

6. The method of claim 1, wherein the semantic model is stored on a user device.

7. The method of claim 1, wherein determining the at least one similarity metric for each document of the plurality of documents includes embedding each document of the plurality of documents in the semantic model.

8. The method of claim 1, wherein computing the at least one relevance metric includes embedding the search query in the semantic model.

9. The method of claim 1, wherein the data store comprises at least a portion of a file system.

10. The method of claim 1, wherein the data store comprises a set of help topic, product or system knowledge base, or other limited domain articles.

11. A system, comprising:

a processor configured to:

provide a data store including a plurality of documents;

provide a semantic model including a plurality of concepts, wherein the semantic model is derived at least in part from a reference source that includes content not included in the data store;

determine at least one similarity metric for each document of the plurality of documents, wherein each respective similarity metric represents a similarity between a respective document of the plurality of documents and a respective concept of the plurality of concepts in the semantic model;

receive a search query;

compute at least one relevance metric of the search query, wherein each relevance metric represents a relevance of the search query to a respective concept of the plurality of concepts represented in the semantic model; and

determine a ranking of at least a subset of the plurality of documents with respect to the search query using at least the at least one similarity metric and the at least one relevance metric; and memory coupled to the processor and configured to store the semantic model.

12. The system of claim 11, wherein the semantic model is stored on the system.

13. The system of claim 11, wherein determining the at least one similarity metric for each document of the plurality of documents includes embedding each document of the plurality of documents in the semantic model.

14. The system of claim 11, wherein computing the at least one relevance metric includes embedding the search query in the semantic model.

15. A computer program product, the computer program product being embodied in a tangible, non-transitory computer readable storage medium and comprising computer instructions for:

providing a data store including a plurality of documents;

providing a semantic model including a plurality of concepts, wherein the semantic model is derived at least in part from a reference source that includes content not included in the data store;

determining at least one similarity metric for each document of the plurality of documents wherein each respective similarity metric represents a similarity between a respective document of the plurality of documents and a respective concept of the plurality of concepts in the semantic model;

receiving a search query;

computing at least one relevance metric of the search query, wherein each relevance metric represents a relevance of the search query to a respective concept of the plurality of concepts represented in the semantic model; and

determining a ranking of at least a subset of the plurality of documents with respect to the search query using at least the at least one similarity metric and the at least one relevance metric.

16. The computer program product of claim 15, further comprising computer instructions for using the reference source to build the semantic model.

17. The computer program product of claim 15, wherein determining the at least one similarity metric for each document of the plurality of documents includes embedding each document of the plurality of documents in the semantic model.

18. The method of claim 5, wherein the reference source is Wikipedia.